Btrfs: fix file corruption after snapshotting due to mix of buffered/DIO writes
[sfrench/cifs-2.6.git] / fs / btrfs / transaction.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5
6 #include <linux/fs.h>
7 #include <linux/slab.h>
8 #include <linux/sched.h>
9 #include <linux/writeback.h>
10 #include <linux/pagemap.h>
11 #include <linux/blkdev.h>
12 #include <linux/uuid.h>
13 #include "ctree.h"
14 #include "disk-io.h"
15 #include "transaction.h"
16 #include "locking.h"
17 #include "tree-log.h"
18 #include "inode-map.h"
19 #include "volumes.h"
20 #include "dev-replace.h"
21 #include "qgroup.h"
22
23 #define BTRFS_ROOT_TRANS_TAG 0
24
25 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
26         [TRANS_STATE_RUNNING]           = 0U,
27         [TRANS_STATE_BLOCKED]           =  __TRANS_START,
28         [TRANS_STATE_COMMIT_START]      = (__TRANS_START | __TRANS_ATTACH),
29         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_START |
30                                            __TRANS_ATTACH |
31                                            __TRANS_JOIN),
32         [TRANS_STATE_UNBLOCKED]         = (__TRANS_START |
33                                            __TRANS_ATTACH |
34                                            __TRANS_JOIN |
35                                            __TRANS_JOIN_NOLOCK),
36         [TRANS_STATE_COMPLETED]         = (__TRANS_START |
37                                            __TRANS_ATTACH |
38                                            __TRANS_JOIN |
39                                            __TRANS_JOIN_NOLOCK),
40 };
41
42 void btrfs_put_transaction(struct btrfs_transaction *transaction)
43 {
44         WARN_ON(refcount_read(&transaction->use_count) == 0);
45         if (refcount_dec_and_test(&transaction->use_count)) {
46                 BUG_ON(!list_empty(&transaction->list));
47                 WARN_ON(!RB_EMPTY_ROOT(
48                                 &transaction->delayed_refs.href_root.rb_root));
49                 if (transaction->delayed_refs.pending_csums)
50                         btrfs_err(transaction->fs_info,
51                                   "pending csums is %llu",
52                                   transaction->delayed_refs.pending_csums);
53                 while (!list_empty(&transaction->pending_chunks)) {
54                         struct extent_map *em;
55
56                         em = list_first_entry(&transaction->pending_chunks,
57                                               struct extent_map, list);
58                         list_del_init(&em->list);
59                         free_extent_map(em);
60                 }
61                 /*
62                  * If any block groups are found in ->deleted_bgs then it's
63                  * because the transaction was aborted and a commit did not
64                  * happen (things failed before writing the new superblock
65                  * and calling btrfs_finish_extent_commit()), so we can not
66                  * discard the physical locations of the block groups.
67                  */
68                 while (!list_empty(&transaction->deleted_bgs)) {
69                         struct btrfs_block_group_cache *cache;
70
71                         cache = list_first_entry(&transaction->deleted_bgs,
72                                                  struct btrfs_block_group_cache,
73                                                  bg_list);
74                         list_del_init(&cache->bg_list);
75                         btrfs_put_block_group_trimming(cache);
76                         btrfs_put_block_group(cache);
77                 }
78                 kfree(transaction);
79         }
80 }
81
82 static void clear_btree_io_tree(struct extent_io_tree *tree)
83 {
84         spin_lock(&tree->lock);
85         /*
86          * Do a single barrier for the waitqueue_active check here, the state
87          * of the waitqueue should not change once clear_btree_io_tree is
88          * called.
89          */
90         smp_mb();
91         while (!RB_EMPTY_ROOT(&tree->state)) {
92                 struct rb_node *node;
93                 struct extent_state *state;
94
95                 node = rb_first(&tree->state);
96                 state = rb_entry(node, struct extent_state, rb_node);
97                 rb_erase(&state->rb_node, &tree->state);
98                 RB_CLEAR_NODE(&state->rb_node);
99                 /*
100                  * btree io trees aren't supposed to have tasks waiting for
101                  * changes in the flags of extent states ever.
102                  */
103                 ASSERT(!waitqueue_active(&state->wq));
104                 free_extent_state(state);
105
106                 cond_resched_lock(&tree->lock);
107         }
108         spin_unlock(&tree->lock);
109 }
110
111 static noinline void switch_commit_roots(struct btrfs_transaction *trans)
112 {
113         struct btrfs_fs_info *fs_info = trans->fs_info;
114         struct btrfs_root *root, *tmp;
115
116         down_write(&fs_info->commit_root_sem);
117         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
118                                  dirty_list) {
119                 list_del_init(&root->dirty_list);
120                 free_extent_buffer(root->commit_root);
121                 root->commit_root = btrfs_root_node(root);
122                 if (is_fstree(root->root_key.objectid))
123                         btrfs_unpin_free_ino(root);
124                 clear_btree_io_tree(&root->dirty_log_pages);
125                 btrfs_qgroup_clean_swapped_blocks(root);
126         }
127
128         /* We can free old roots now. */
129         spin_lock(&trans->dropped_roots_lock);
130         while (!list_empty(&trans->dropped_roots)) {
131                 root = list_first_entry(&trans->dropped_roots,
132                                         struct btrfs_root, root_list);
133                 list_del_init(&root->root_list);
134                 spin_unlock(&trans->dropped_roots_lock);
135                 btrfs_drop_and_free_fs_root(fs_info, root);
136                 spin_lock(&trans->dropped_roots_lock);
137         }
138         spin_unlock(&trans->dropped_roots_lock);
139         up_write(&fs_info->commit_root_sem);
140 }
141
142 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
143                                          unsigned int type)
144 {
145         if (type & TRANS_EXTWRITERS)
146                 atomic_inc(&trans->num_extwriters);
147 }
148
149 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
150                                          unsigned int type)
151 {
152         if (type & TRANS_EXTWRITERS)
153                 atomic_dec(&trans->num_extwriters);
154 }
155
156 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
157                                           unsigned int type)
158 {
159         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
160 }
161
162 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
163 {
164         return atomic_read(&trans->num_extwriters);
165 }
166
167 /*
168  * either allocate a new transaction or hop into the existing one
169  */
170 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
171                                      unsigned int type)
172 {
173         struct btrfs_transaction *cur_trans;
174
175         spin_lock(&fs_info->trans_lock);
176 loop:
177         /* The file system has been taken offline. No new transactions. */
178         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
179                 spin_unlock(&fs_info->trans_lock);
180                 return -EROFS;
181         }
182
183         cur_trans = fs_info->running_transaction;
184         if (cur_trans) {
185                 if (cur_trans->aborted) {
186                         spin_unlock(&fs_info->trans_lock);
187                         return cur_trans->aborted;
188                 }
189                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
190                         spin_unlock(&fs_info->trans_lock);
191                         return -EBUSY;
192                 }
193                 refcount_inc(&cur_trans->use_count);
194                 atomic_inc(&cur_trans->num_writers);
195                 extwriter_counter_inc(cur_trans, type);
196                 spin_unlock(&fs_info->trans_lock);
197                 return 0;
198         }
199         spin_unlock(&fs_info->trans_lock);
200
201         /*
202          * If we are ATTACH, we just want to catch the current transaction,
203          * and commit it. If there is no transaction, just return ENOENT.
204          */
205         if (type == TRANS_ATTACH)
206                 return -ENOENT;
207
208         /*
209          * JOIN_NOLOCK only happens during the transaction commit, so
210          * it is impossible that ->running_transaction is NULL
211          */
212         BUG_ON(type == TRANS_JOIN_NOLOCK);
213
214         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
215         if (!cur_trans)
216                 return -ENOMEM;
217
218         spin_lock(&fs_info->trans_lock);
219         if (fs_info->running_transaction) {
220                 /*
221                  * someone started a transaction after we unlocked.  Make sure
222                  * to redo the checks above
223                  */
224                 kfree(cur_trans);
225                 goto loop;
226         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
227                 spin_unlock(&fs_info->trans_lock);
228                 kfree(cur_trans);
229                 return -EROFS;
230         }
231
232         cur_trans->fs_info = fs_info;
233         atomic_set(&cur_trans->num_writers, 1);
234         extwriter_counter_init(cur_trans, type);
235         init_waitqueue_head(&cur_trans->writer_wait);
236         init_waitqueue_head(&cur_trans->commit_wait);
237         cur_trans->state = TRANS_STATE_RUNNING;
238         /*
239          * One for this trans handle, one so it will live on until we
240          * commit the transaction.
241          */
242         refcount_set(&cur_trans->use_count, 2);
243         cur_trans->flags = 0;
244         cur_trans->start_time = ktime_get_seconds();
245
246         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
247
248         cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
249         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
250         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
251
252         /*
253          * although the tree mod log is per file system and not per transaction,
254          * the log must never go across transaction boundaries.
255          */
256         smp_mb();
257         if (!list_empty(&fs_info->tree_mod_seq_list))
258                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
259         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
260                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
261         atomic64_set(&fs_info->tree_mod_seq, 0);
262
263         spin_lock_init(&cur_trans->delayed_refs.lock);
264
265         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
266         INIT_LIST_HEAD(&cur_trans->pending_chunks);
267         INIT_LIST_HEAD(&cur_trans->switch_commits);
268         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
269         INIT_LIST_HEAD(&cur_trans->io_bgs);
270         INIT_LIST_HEAD(&cur_trans->dropped_roots);
271         mutex_init(&cur_trans->cache_write_mutex);
272         cur_trans->num_dirty_bgs = 0;
273         spin_lock_init(&cur_trans->dirty_bgs_lock);
274         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
275         spin_lock_init(&cur_trans->dropped_roots_lock);
276         list_add_tail(&cur_trans->list, &fs_info->trans_list);
277         extent_io_tree_init(&cur_trans->dirty_pages,
278                              fs_info->btree_inode);
279         fs_info->generation++;
280         cur_trans->transid = fs_info->generation;
281         fs_info->running_transaction = cur_trans;
282         cur_trans->aborted = 0;
283         spin_unlock(&fs_info->trans_lock);
284
285         return 0;
286 }
287
288 /*
289  * this does all the record keeping required to make sure that a reference
290  * counted root is properly recorded in a given transaction.  This is required
291  * to make sure the old root from before we joined the transaction is deleted
292  * when the transaction commits
293  */
294 static int record_root_in_trans(struct btrfs_trans_handle *trans,
295                                struct btrfs_root *root,
296                                int force)
297 {
298         struct btrfs_fs_info *fs_info = root->fs_info;
299
300         if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
301             root->last_trans < trans->transid) || force) {
302                 WARN_ON(root == fs_info->extent_root);
303                 WARN_ON(!force && root->commit_root != root->node);
304
305                 /*
306                  * see below for IN_TRANS_SETUP usage rules
307                  * we have the reloc mutex held now, so there
308                  * is only one writer in this function
309                  */
310                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
311
312                 /* make sure readers find IN_TRANS_SETUP before
313                  * they find our root->last_trans update
314                  */
315                 smp_wmb();
316
317                 spin_lock(&fs_info->fs_roots_radix_lock);
318                 if (root->last_trans == trans->transid && !force) {
319                         spin_unlock(&fs_info->fs_roots_radix_lock);
320                         return 0;
321                 }
322                 radix_tree_tag_set(&fs_info->fs_roots_radix,
323                                    (unsigned long)root->root_key.objectid,
324                                    BTRFS_ROOT_TRANS_TAG);
325                 spin_unlock(&fs_info->fs_roots_radix_lock);
326                 root->last_trans = trans->transid;
327
328                 /* this is pretty tricky.  We don't want to
329                  * take the relocation lock in btrfs_record_root_in_trans
330                  * unless we're really doing the first setup for this root in
331                  * this transaction.
332                  *
333                  * Normally we'd use root->last_trans as a flag to decide
334                  * if we want to take the expensive mutex.
335                  *
336                  * But, we have to set root->last_trans before we
337                  * init the relocation root, otherwise, we trip over warnings
338                  * in ctree.c.  The solution used here is to flag ourselves
339                  * with root IN_TRANS_SETUP.  When this is 1, we're still
340                  * fixing up the reloc trees and everyone must wait.
341                  *
342                  * When this is zero, they can trust root->last_trans and fly
343                  * through btrfs_record_root_in_trans without having to take the
344                  * lock.  smp_wmb() makes sure that all the writes above are
345                  * done before we pop in the zero below
346                  */
347                 btrfs_init_reloc_root(trans, root);
348                 smp_mb__before_atomic();
349                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
350         }
351         return 0;
352 }
353
354
355 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
356                             struct btrfs_root *root)
357 {
358         struct btrfs_fs_info *fs_info = root->fs_info;
359         struct btrfs_transaction *cur_trans = trans->transaction;
360
361         /* Add ourselves to the transaction dropped list */
362         spin_lock(&cur_trans->dropped_roots_lock);
363         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
364         spin_unlock(&cur_trans->dropped_roots_lock);
365
366         /* Make sure we don't try to update the root at commit time */
367         spin_lock(&fs_info->fs_roots_radix_lock);
368         radix_tree_tag_clear(&fs_info->fs_roots_radix,
369                              (unsigned long)root->root_key.objectid,
370                              BTRFS_ROOT_TRANS_TAG);
371         spin_unlock(&fs_info->fs_roots_radix_lock);
372 }
373
374 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
375                                struct btrfs_root *root)
376 {
377         struct btrfs_fs_info *fs_info = root->fs_info;
378
379         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
380                 return 0;
381
382         /*
383          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
384          * and barriers
385          */
386         smp_rmb();
387         if (root->last_trans == trans->transid &&
388             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
389                 return 0;
390
391         mutex_lock(&fs_info->reloc_mutex);
392         record_root_in_trans(trans, root, 0);
393         mutex_unlock(&fs_info->reloc_mutex);
394
395         return 0;
396 }
397
398 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
399 {
400         return (trans->state >= TRANS_STATE_BLOCKED &&
401                 trans->state < TRANS_STATE_UNBLOCKED &&
402                 !trans->aborted);
403 }
404
405 /* wait for commit against the current transaction to become unblocked
406  * when this is done, it is safe to start a new transaction, but the current
407  * transaction might not be fully on disk.
408  */
409 static void wait_current_trans(struct btrfs_fs_info *fs_info)
410 {
411         struct btrfs_transaction *cur_trans;
412
413         spin_lock(&fs_info->trans_lock);
414         cur_trans = fs_info->running_transaction;
415         if (cur_trans && is_transaction_blocked(cur_trans)) {
416                 refcount_inc(&cur_trans->use_count);
417                 spin_unlock(&fs_info->trans_lock);
418
419                 wait_event(fs_info->transaction_wait,
420                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
421                            cur_trans->aborted);
422                 btrfs_put_transaction(cur_trans);
423         } else {
424                 spin_unlock(&fs_info->trans_lock);
425         }
426 }
427
428 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
429 {
430         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
431                 return 0;
432
433         if (type == TRANS_START)
434                 return 1;
435
436         return 0;
437 }
438
439 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
440 {
441         struct btrfs_fs_info *fs_info = root->fs_info;
442
443         if (!fs_info->reloc_ctl ||
444             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
445             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
446             root->reloc_root)
447                 return false;
448
449         return true;
450 }
451
452 static struct btrfs_trans_handle *
453 start_transaction(struct btrfs_root *root, unsigned int num_items,
454                   unsigned int type, enum btrfs_reserve_flush_enum flush,
455                   bool enforce_qgroups)
456 {
457         struct btrfs_fs_info *fs_info = root->fs_info;
458         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
459         struct btrfs_trans_handle *h;
460         struct btrfs_transaction *cur_trans;
461         u64 num_bytes = 0;
462         u64 qgroup_reserved = 0;
463         bool reloc_reserved = false;
464         int ret;
465
466         /* Send isn't supposed to start transactions. */
467         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
468
469         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
470                 return ERR_PTR(-EROFS);
471
472         if (current->journal_info) {
473                 WARN_ON(type & TRANS_EXTWRITERS);
474                 h = current->journal_info;
475                 refcount_inc(&h->use_count);
476                 WARN_ON(refcount_read(&h->use_count) > 2);
477                 h->orig_rsv = h->block_rsv;
478                 h->block_rsv = NULL;
479                 goto got_it;
480         }
481
482         /*
483          * Do the reservation before we join the transaction so we can do all
484          * the appropriate flushing if need be.
485          */
486         if (num_items && root != fs_info->chunk_root) {
487                 struct btrfs_block_rsv *rsv = &fs_info->trans_block_rsv;
488                 u64 delayed_refs_bytes = 0;
489
490                 qgroup_reserved = num_items * fs_info->nodesize;
491                 ret = btrfs_qgroup_reserve_meta_pertrans(root, qgroup_reserved,
492                                 enforce_qgroups);
493                 if (ret)
494                         return ERR_PTR(ret);
495
496                 /*
497                  * We want to reserve all the bytes we may need all at once, so
498                  * we only do 1 enospc flushing cycle per transaction start.  We
499                  * accomplish this by simply assuming we'll do 2 x num_items
500                  * worth of delayed refs updates in this trans handle, and
501                  * refill that amount for whatever is missing in the reserve.
502                  */
503                 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
504                 if (delayed_refs_rsv->full == 0) {
505                         delayed_refs_bytes = num_bytes;
506                         num_bytes <<= 1;
507                 }
508
509                 /*
510                  * Do the reservation for the relocation root creation
511                  */
512                 if (need_reserve_reloc_root(root)) {
513                         num_bytes += fs_info->nodesize;
514                         reloc_reserved = true;
515                 }
516
517                 ret = btrfs_block_rsv_add(root, rsv, num_bytes, flush);
518                 if (ret)
519                         goto reserve_fail;
520                 if (delayed_refs_bytes) {
521                         btrfs_migrate_to_delayed_refs_rsv(fs_info, rsv,
522                                                           delayed_refs_bytes);
523                         num_bytes -= delayed_refs_bytes;
524                 }
525         } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
526                    !delayed_refs_rsv->full) {
527                 /*
528                  * Some people call with btrfs_start_transaction(root, 0)
529                  * because they can be throttled, but have some other mechanism
530                  * for reserving space.  We still want these guys to refill the
531                  * delayed block_rsv so just add 1 items worth of reservation
532                  * here.
533                  */
534                 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
535                 if (ret)
536                         goto reserve_fail;
537         }
538 again:
539         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
540         if (!h) {
541                 ret = -ENOMEM;
542                 goto alloc_fail;
543         }
544
545         /*
546          * If we are JOIN_NOLOCK we're already committing a transaction and
547          * waiting on this guy, so we don't need to do the sb_start_intwrite
548          * because we're already holding a ref.  We need this because we could
549          * have raced in and did an fsync() on a file which can kick a commit
550          * and then we deadlock with somebody doing a freeze.
551          *
552          * If we are ATTACH, it means we just want to catch the current
553          * transaction and commit it, so we needn't do sb_start_intwrite(). 
554          */
555         if (type & __TRANS_FREEZABLE)
556                 sb_start_intwrite(fs_info->sb);
557
558         if (may_wait_transaction(fs_info, type))
559                 wait_current_trans(fs_info);
560
561         do {
562                 ret = join_transaction(fs_info, type);
563                 if (ret == -EBUSY) {
564                         wait_current_trans(fs_info);
565                         if (unlikely(type == TRANS_ATTACH))
566                                 ret = -ENOENT;
567                 }
568         } while (ret == -EBUSY);
569
570         if (ret < 0)
571                 goto join_fail;
572
573         cur_trans = fs_info->running_transaction;
574
575         h->transid = cur_trans->transid;
576         h->transaction = cur_trans;
577         h->root = root;
578         refcount_set(&h->use_count, 1);
579         h->fs_info = root->fs_info;
580
581         h->type = type;
582         h->can_flush_pending_bgs = true;
583         INIT_LIST_HEAD(&h->new_bgs);
584
585         smp_mb();
586         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
587             may_wait_transaction(fs_info, type)) {
588                 current->journal_info = h;
589                 btrfs_commit_transaction(h);
590                 goto again;
591         }
592
593         if (num_bytes) {
594                 trace_btrfs_space_reservation(fs_info, "transaction",
595                                               h->transid, num_bytes, 1);
596                 h->block_rsv = &fs_info->trans_block_rsv;
597                 h->bytes_reserved = num_bytes;
598                 h->reloc_reserved = reloc_reserved;
599         }
600
601 got_it:
602         btrfs_record_root_in_trans(h, root);
603
604         if (!current->journal_info)
605                 current->journal_info = h;
606         return h;
607
608 join_fail:
609         if (type & __TRANS_FREEZABLE)
610                 sb_end_intwrite(fs_info->sb);
611         kmem_cache_free(btrfs_trans_handle_cachep, h);
612 alloc_fail:
613         if (num_bytes)
614                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
615                                         num_bytes);
616 reserve_fail:
617         btrfs_qgroup_free_meta_pertrans(root, qgroup_reserved);
618         return ERR_PTR(ret);
619 }
620
621 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
622                                                    unsigned int num_items)
623 {
624         return start_transaction(root, num_items, TRANS_START,
625                                  BTRFS_RESERVE_FLUSH_ALL, true);
626 }
627
628 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
629                                         struct btrfs_root *root,
630                                         unsigned int num_items,
631                                         int min_factor)
632 {
633         struct btrfs_fs_info *fs_info = root->fs_info;
634         struct btrfs_trans_handle *trans;
635         u64 num_bytes;
636         int ret;
637
638         /*
639          * We have two callers: unlink and block group removal.  The
640          * former should succeed even if we will temporarily exceed
641          * quota and the latter operates on the extent root so
642          * qgroup enforcement is ignored anyway.
643          */
644         trans = start_transaction(root, num_items, TRANS_START,
645                                   BTRFS_RESERVE_FLUSH_ALL, false);
646         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
647                 return trans;
648
649         trans = btrfs_start_transaction(root, 0);
650         if (IS_ERR(trans))
651                 return trans;
652
653         num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
654         ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
655                                        num_bytes, min_factor);
656         if (ret) {
657                 btrfs_end_transaction(trans);
658                 return ERR_PTR(ret);
659         }
660
661         trans->block_rsv = &fs_info->trans_block_rsv;
662         trans->bytes_reserved = num_bytes;
663         trace_btrfs_space_reservation(fs_info, "transaction",
664                                       trans->transid, num_bytes, 1);
665
666         return trans;
667 }
668
669 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
670 {
671         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
672                                  true);
673 }
674
675 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
676 {
677         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
678                                  BTRFS_RESERVE_NO_FLUSH, true);
679 }
680
681 /*
682  * btrfs_attach_transaction() - catch the running transaction
683  *
684  * It is used when we want to commit the current the transaction, but
685  * don't want to start a new one.
686  *
687  * Note: If this function return -ENOENT, it just means there is no
688  * running transaction. But it is possible that the inactive transaction
689  * is still in the memory, not fully on disk. If you hope there is no
690  * inactive transaction in the fs when -ENOENT is returned, you should
691  * invoke
692  *     btrfs_attach_transaction_barrier()
693  */
694 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
695 {
696         return start_transaction(root, 0, TRANS_ATTACH,
697                                  BTRFS_RESERVE_NO_FLUSH, true);
698 }
699
700 /*
701  * btrfs_attach_transaction_barrier() - catch the running transaction
702  *
703  * It is similar to the above function, the difference is this one
704  * will wait for all the inactive transactions until they fully
705  * complete.
706  */
707 struct btrfs_trans_handle *
708 btrfs_attach_transaction_barrier(struct btrfs_root *root)
709 {
710         struct btrfs_trans_handle *trans;
711
712         trans = start_transaction(root, 0, TRANS_ATTACH,
713                                   BTRFS_RESERVE_NO_FLUSH, true);
714         if (trans == ERR_PTR(-ENOENT))
715                 btrfs_wait_for_commit(root->fs_info, 0);
716
717         return trans;
718 }
719
720 /* wait for a transaction commit to be fully complete */
721 static noinline void wait_for_commit(struct btrfs_transaction *commit)
722 {
723         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
724 }
725
726 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
727 {
728         struct btrfs_transaction *cur_trans = NULL, *t;
729         int ret = 0;
730
731         if (transid) {
732                 if (transid <= fs_info->last_trans_committed)
733                         goto out;
734
735                 /* find specified transaction */
736                 spin_lock(&fs_info->trans_lock);
737                 list_for_each_entry(t, &fs_info->trans_list, list) {
738                         if (t->transid == transid) {
739                                 cur_trans = t;
740                                 refcount_inc(&cur_trans->use_count);
741                                 ret = 0;
742                                 break;
743                         }
744                         if (t->transid > transid) {
745                                 ret = 0;
746                                 break;
747                         }
748                 }
749                 spin_unlock(&fs_info->trans_lock);
750
751                 /*
752                  * The specified transaction doesn't exist, or we
753                  * raced with btrfs_commit_transaction
754                  */
755                 if (!cur_trans) {
756                         if (transid > fs_info->last_trans_committed)
757                                 ret = -EINVAL;
758                         goto out;
759                 }
760         } else {
761                 /* find newest transaction that is committing | committed */
762                 spin_lock(&fs_info->trans_lock);
763                 list_for_each_entry_reverse(t, &fs_info->trans_list,
764                                             list) {
765                         if (t->state >= TRANS_STATE_COMMIT_START) {
766                                 if (t->state == TRANS_STATE_COMPLETED)
767                                         break;
768                                 cur_trans = t;
769                                 refcount_inc(&cur_trans->use_count);
770                                 break;
771                         }
772                 }
773                 spin_unlock(&fs_info->trans_lock);
774                 if (!cur_trans)
775                         goto out;  /* nothing committing|committed */
776         }
777
778         wait_for_commit(cur_trans);
779         btrfs_put_transaction(cur_trans);
780 out:
781         return ret;
782 }
783
784 void btrfs_throttle(struct btrfs_fs_info *fs_info)
785 {
786         wait_current_trans(fs_info);
787 }
788
789 static int should_end_transaction(struct btrfs_trans_handle *trans)
790 {
791         struct btrfs_fs_info *fs_info = trans->fs_info;
792
793         if (btrfs_check_space_for_delayed_refs(fs_info))
794                 return 1;
795
796         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
797 }
798
799 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
800 {
801         struct btrfs_transaction *cur_trans = trans->transaction;
802
803         smp_mb();
804         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
805             cur_trans->delayed_refs.flushing)
806                 return 1;
807
808         return should_end_transaction(trans);
809 }
810
811 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
812
813 {
814         struct btrfs_fs_info *fs_info = trans->fs_info;
815
816         if (!trans->block_rsv) {
817                 ASSERT(!trans->bytes_reserved);
818                 return;
819         }
820
821         if (!trans->bytes_reserved)
822                 return;
823
824         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
825         trace_btrfs_space_reservation(fs_info, "transaction",
826                                       trans->transid, trans->bytes_reserved, 0);
827         btrfs_block_rsv_release(fs_info, trans->block_rsv,
828                                 trans->bytes_reserved);
829         trans->bytes_reserved = 0;
830 }
831
832 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
833                                    int throttle)
834 {
835         struct btrfs_fs_info *info = trans->fs_info;
836         struct btrfs_transaction *cur_trans = trans->transaction;
837         int lock = (trans->type != TRANS_JOIN_NOLOCK);
838         int err = 0;
839
840         if (refcount_read(&trans->use_count) > 1) {
841                 refcount_dec(&trans->use_count);
842                 trans->block_rsv = trans->orig_rsv;
843                 return 0;
844         }
845
846         btrfs_trans_release_metadata(trans);
847         trans->block_rsv = NULL;
848
849         btrfs_create_pending_block_groups(trans);
850
851         btrfs_trans_release_chunk_metadata(trans);
852
853         if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
854                 if (throttle)
855                         return btrfs_commit_transaction(trans);
856                 else
857                         wake_up_process(info->transaction_kthread);
858         }
859
860         if (trans->type & __TRANS_FREEZABLE)
861                 sb_end_intwrite(info->sb);
862
863         WARN_ON(cur_trans != info->running_transaction);
864         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
865         atomic_dec(&cur_trans->num_writers);
866         extwriter_counter_dec(cur_trans, trans->type);
867
868         cond_wake_up(&cur_trans->writer_wait);
869         btrfs_put_transaction(cur_trans);
870
871         if (current->journal_info == trans)
872                 current->journal_info = NULL;
873
874         if (throttle)
875                 btrfs_run_delayed_iputs(info);
876
877         if (trans->aborted ||
878             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
879                 wake_up_process(info->transaction_kthread);
880                 err = -EIO;
881         }
882
883         kmem_cache_free(btrfs_trans_handle_cachep, trans);
884         return err;
885 }
886
887 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
888 {
889         return __btrfs_end_transaction(trans, 0);
890 }
891
892 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
893 {
894         return __btrfs_end_transaction(trans, 1);
895 }
896
897 /*
898  * when btree blocks are allocated, they have some corresponding bits set for
899  * them in one of two extent_io trees.  This is used to make sure all of
900  * those extents are sent to disk but does not wait on them
901  */
902 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
903                                struct extent_io_tree *dirty_pages, int mark)
904 {
905         int err = 0;
906         int werr = 0;
907         struct address_space *mapping = fs_info->btree_inode->i_mapping;
908         struct extent_state *cached_state = NULL;
909         u64 start = 0;
910         u64 end;
911
912         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
913         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
914                                       mark, &cached_state)) {
915                 bool wait_writeback = false;
916
917                 err = convert_extent_bit(dirty_pages, start, end,
918                                          EXTENT_NEED_WAIT,
919                                          mark, &cached_state);
920                 /*
921                  * convert_extent_bit can return -ENOMEM, which is most of the
922                  * time a temporary error. So when it happens, ignore the error
923                  * and wait for writeback of this range to finish - because we
924                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
925                  * to __btrfs_wait_marked_extents() would not know that
926                  * writeback for this range started and therefore wouldn't
927                  * wait for it to finish - we don't want to commit a
928                  * superblock that points to btree nodes/leafs for which
929                  * writeback hasn't finished yet (and without errors).
930                  * We cleanup any entries left in the io tree when committing
931                  * the transaction (through clear_btree_io_tree()).
932                  */
933                 if (err == -ENOMEM) {
934                         err = 0;
935                         wait_writeback = true;
936                 }
937                 if (!err)
938                         err = filemap_fdatawrite_range(mapping, start, end);
939                 if (err)
940                         werr = err;
941                 else if (wait_writeback)
942                         werr = filemap_fdatawait_range(mapping, start, end);
943                 free_extent_state(cached_state);
944                 cached_state = NULL;
945                 cond_resched();
946                 start = end + 1;
947         }
948         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
949         return werr;
950 }
951
952 /*
953  * when btree blocks are allocated, they have some corresponding bits set for
954  * them in one of two extent_io trees.  This is used to make sure all of
955  * those extents are on disk for transaction or log commit.  We wait
956  * on all the pages and clear them from the dirty pages state tree
957  */
958 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
959                                        struct extent_io_tree *dirty_pages)
960 {
961         int err = 0;
962         int werr = 0;
963         struct address_space *mapping = fs_info->btree_inode->i_mapping;
964         struct extent_state *cached_state = NULL;
965         u64 start = 0;
966         u64 end;
967
968         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
969                                       EXTENT_NEED_WAIT, &cached_state)) {
970                 /*
971                  * Ignore -ENOMEM errors returned by clear_extent_bit().
972                  * When committing the transaction, we'll remove any entries
973                  * left in the io tree. For a log commit, we don't remove them
974                  * after committing the log because the tree can be accessed
975                  * concurrently - we do it only at transaction commit time when
976                  * it's safe to do it (through clear_btree_io_tree()).
977                  */
978                 err = clear_extent_bit(dirty_pages, start, end,
979                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
980                 if (err == -ENOMEM)
981                         err = 0;
982                 if (!err)
983                         err = filemap_fdatawait_range(mapping, start, end);
984                 if (err)
985                         werr = err;
986                 free_extent_state(cached_state);
987                 cached_state = NULL;
988                 cond_resched();
989                 start = end + 1;
990         }
991         if (err)
992                 werr = err;
993         return werr;
994 }
995
996 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
997                        struct extent_io_tree *dirty_pages)
998 {
999         bool errors = false;
1000         int err;
1001
1002         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1003         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1004                 errors = true;
1005
1006         if (errors && !err)
1007                 err = -EIO;
1008         return err;
1009 }
1010
1011 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1012 {
1013         struct btrfs_fs_info *fs_info = log_root->fs_info;
1014         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1015         bool errors = false;
1016         int err;
1017
1018         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1019
1020         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1021         if ((mark & EXTENT_DIRTY) &&
1022             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1023                 errors = true;
1024
1025         if ((mark & EXTENT_NEW) &&
1026             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1027                 errors = true;
1028
1029         if (errors && !err)
1030                 err = -EIO;
1031         return err;
1032 }
1033
1034 /*
1035  * When btree blocks are allocated the corresponding extents are marked dirty.
1036  * This function ensures such extents are persisted on disk for transaction or
1037  * log commit.
1038  *
1039  * @trans: transaction whose dirty pages we'd like to write
1040  */
1041 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1042 {
1043         int ret;
1044         int ret2;
1045         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1046         struct btrfs_fs_info *fs_info = trans->fs_info;
1047         struct blk_plug plug;
1048
1049         blk_start_plug(&plug);
1050         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1051         blk_finish_plug(&plug);
1052         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1053
1054         clear_btree_io_tree(&trans->transaction->dirty_pages);
1055
1056         if (ret)
1057                 return ret;
1058         else if (ret2)
1059                 return ret2;
1060         else
1061                 return 0;
1062 }
1063
1064 /*
1065  * this is used to update the root pointer in the tree of tree roots.
1066  *
1067  * But, in the case of the extent allocation tree, updating the root
1068  * pointer may allocate blocks which may change the root of the extent
1069  * allocation tree.
1070  *
1071  * So, this loops and repeats and makes sure the cowonly root didn't
1072  * change while the root pointer was being updated in the metadata.
1073  */
1074 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1075                                struct btrfs_root *root)
1076 {
1077         int ret;
1078         u64 old_root_bytenr;
1079         u64 old_root_used;
1080         struct btrfs_fs_info *fs_info = root->fs_info;
1081         struct btrfs_root *tree_root = fs_info->tree_root;
1082
1083         old_root_used = btrfs_root_used(&root->root_item);
1084
1085         while (1) {
1086                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1087                 if (old_root_bytenr == root->node->start &&
1088                     old_root_used == btrfs_root_used(&root->root_item))
1089                         break;
1090
1091                 btrfs_set_root_node(&root->root_item, root->node);
1092                 ret = btrfs_update_root(trans, tree_root,
1093                                         &root->root_key,
1094                                         &root->root_item);
1095                 if (ret)
1096                         return ret;
1097
1098                 old_root_used = btrfs_root_used(&root->root_item);
1099         }
1100
1101         return 0;
1102 }
1103
1104 /*
1105  * update all the cowonly tree roots on disk
1106  *
1107  * The error handling in this function may not be obvious. Any of the
1108  * failures will cause the file system to go offline. We still need
1109  * to clean up the delayed refs.
1110  */
1111 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1112 {
1113         struct btrfs_fs_info *fs_info = trans->fs_info;
1114         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1115         struct list_head *io_bgs = &trans->transaction->io_bgs;
1116         struct list_head *next;
1117         struct extent_buffer *eb;
1118         int ret;
1119
1120         eb = btrfs_lock_root_node(fs_info->tree_root);
1121         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1122                               0, &eb);
1123         btrfs_tree_unlock(eb);
1124         free_extent_buffer(eb);
1125
1126         if (ret)
1127                 return ret;
1128
1129         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1130         if (ret)
1131                 return ret;
1132
1133         ret = btrfs_run_dev_stats(trans, fs_info);
1134         if (ret)
1135                 return ret;
1136         ret = btrfs_run_dev_replace(trans, fs_info);
1137         if (ret)
1138                 return ret;
1139         ret = btrfs_run_qgroups(trans);
1140         if (ret)
1141                 return ret;
1142
1143         ret = btrfs_setup_space_cache(trans, fs_info);
1144         if (ret)
1145                 return ret;
1146
1147         /* run_qgroups might have added some more refs */
1148         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1149         if (ret)
1150                 return ret;
1151 again:
1152         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1153                 struct btrfs_root *root;
1154                 next = fs_info->dirty_cowonly_roots.next;
1155                 list_del_init(next);
1156                 root = list_entry(next, struct btrfs_root, dirty_list);
1157                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1158
1159                 if (root != fs_info->extent_root)
1160                         list_add_tail(&root->dirty_list,
1161                                       &trans->transaction->switch_commits);
1162                 ret = update_cowonly_root(trans, root);
1163                 if (ret)
1164                         return ret;
1165                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1166                 if (ret)
1167                         return ret;
1168         }
1169
1170         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1171                 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1172                 if (ret)
1173                         return ret;
1174                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1175                 if (ret)
1176                         return ret;
1177         }
1178
1179         if (!list_empty(&fs_info->dirty_cowonly_roots))
1180                 goto again;
1181
1182         list_add_tail(&fs_info->extent_root->dirty_list,
1183                       &trans->transaction->switch_commits);
1184
1185         /* Update dev-replace pointer once everything is committed */
1186         fs_info->dev_replace.committed_cursor_left =
1187                 fs_info->dev_replace.cursor_left_last_write_of_item;
1188
1189         return 0;
1190 }
1191
1192 /*
1193  * dead roots are old snapshots that need to be deleted.  This allocates
1194  * a dirty root struct and adds it into the list of dead roots that need to
1195  * be deleted
1196  */
1197 void btrfs_add_dead_root(struct btrfs_root *root)
1198 {
1199         struct btrfs_fs_info *fs_info = root->fs_info;
1200
1201         spin_lock(&fs_info->trans_lock);
1202         if (list_empty(&root->root_list))
1203                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1204         spin_unlock(&fs_info->trans_lock);
1205 }
1206
1207 /*
1208  * update all the cowonly tree roots on disk
1209  */
1210 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1211 {
1212         struct btrfs_fs_info *fs_info = trans->fs_info;
1213         struct btrfs_root *gang[8];
1214         int i;
1215         int ret;
1216         int err = 0;
1217
1218         spin_lock(&fs_info->fs_roots_radix_lock);
1219         while (1) {
1220                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1221                                                  (void **)gang, 0,
1222                                                  ARRAY_SIZE(gang),
1223                                                  BTRFS_ROOT_TRANS_TAG);
1224                 if (ret == 0)
1225                         break;
1226                 for (i = 0; i < ret; i++) {
1227                         struct btrfs_root *root = gang[i];
1228                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1229                                         (unsigned long)root->root_key.objectid,
1230                                         BTRFS_ROOT_TRANS_TAG);
1231                         spin_unlock(&fs_info->fs_roots_radix_lock);
1232
1233                         btrfs_free_log(trans, root);
1234                         btrfs_update_reloc_root(trans, root);
1235
1236                         btrfs_save_ino_cache(root, trans);
1237
1238                         /* see comments in should_cow_block() */
1239                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1240                         smp_mb__after_atomic();
1241
1242                         if (root->commit_root != root->node) {
1243                                 list_add_tail(&root->dirty_list,
1244                                         &trans->transaction->switch_commits);
1245                                 btrfs_set_root_node(&root->root_item,
1246                                                     root->node);
1247                         }
1248
1249                         err = btrfs_update_root(trans, fs_info->tree_root,
1250                                                 &root->root_key,
1251                                                 &root->root_item);
1252                         spin_lock(&fs_info->fs_roots_radix_lock);
1253                         if (err)
1254                                 break;
1255                         btrfs_qgroup_free_meta_all_pertrans(root);
1256                 }
1257         }
1258         spin_unlock(&fs_info->fs_roots_radix_lock);
1259         return err;
1260 }
1261
1262 /*
1263  * defrag a given btree.
1264  * Every leaf in the btree is read and defragged.
1265  */
1266 int btrfs_defrag_root(struct btrfs_root *root)
1267 {
1268         struct btrfs_fs_info *info = root->fs_info;
1269         struct btrfs_trans_handle *trans;
1270         int ret;
1271
1272         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1273                 return 0;
1274
1275         while (1) {
1276                 trans = btrfs_start_transaction(root, 0);
1277                 if (IS_ERR(trans))
1278                         return PTR_ERR(trans);
1279
1280                 ret = btrfs_defrag_leaves(trans, root);
1281
1282                 btrfs_end_transaction(trans);
1283                 btrfs_btree_balance_dirty(info);
1284                 cond_resched();
1285
1286                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1287                         break;
1288
1289                 if (btrfs_defrag_cancelled(info)) {
1290                         btrfs_debug(info, "defrag_root cancelled");
1291                         ret = -EAGAIN;
1292                         break;
1293                 }
1294         }
1295         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1296         return ret;
1297 }
1298
1299 /*
1300  * Do all special snapshot related qgroup dirty hack.
1301  *
1302  * Will do all needed qgroup inherit and dirty hack like switch commit
1303  * roots inside one transaction and write all btree into disk, to make
1304  * qgroup works.
1305  */
1306 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1307                                    struct btrfs_root *src,
1308                                    struct btrfs_root *parent,
1309                                    struct btrfs_qgroup_inherit *inherit,
1310                                    u64 dst_objectid)
1311 {
1312         struct btrfs_fs_info *fs_info = src->fs_info;
1313         int ret;
1314
1315         /*
1316          * Save some performance in the case that qgroups are not
1317          * enabled. If this check races with the ioctl, rescan will
1318          * kick in anyway.
1319          */
1320         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1321                 return 0;
1322
1323         /*
1324          * Ensure dirty @src will be committed.  Or, after coming
1325          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1326          * recorded root will never be updated again, causing an outdated root
1327          * item.
1328          */
1329         record_root_in_trans(trans, src, 1);
1330
1331         /*
1332          * We are going to commit transaction, see btrfs_commit_transaction()
1333          * comment for reason locking tree_log_mutex
1334          */
1335         mutex_lock(&fs_info->tree_log_mutex);
1336
1337         ret = commit_fs_roots(trans);
1338         if (ret)
1339                 goto out;
1340         ret = btrfs_qgroup_account_extents(trans);
1341         if (ret < 0)
1342                 goto out;
1343
1344         /* Now qgroup are all updated, we can inherit it to new qgroups */
1345         ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1346                                    inherit);
1347         if (ret < 0)
1348                 goto out;
1349
1350         /*
1351          * Now we do a simplified commit transaction, which will:
1352          * 1) commit all subvolume and extent tree
1353          *    To ensure all subvolume and extent tree have a valid
1354          *    commit_root to accounting later insert_dir_item()
1355          * 2) write all btree blocks onto disk
1356          *    This is to make sure later btree modification will be cowed
1357          *    Or commit_root can be populated and cause wrong qgroup numbers
1358          * In this simplified commit, we don't really care about other trees
1359          * like chunk and root tree, as they won't affect qgroup.
1360          * And we don't write super to avoid half committed status.
1361          */
1362         ret = commit_cowonly_roots(trans);
1363         if (ret)
1364                 goto out;
1365         switch_commit_roots(trans->transaction);
1366         ret = btrfs_write_and_wait_transaction(trans);
1367         if (ret)
1368                 btrfs_handle_fs_error(fs_info, ret,
1369                         "Error while writing out transaction for qgroup");
1370
1371 out:
1372         mutex_unlock(&fs_info->tree_log_mutex);
1373
1374         /*
1375          * Force parent root to be updated, as we recorded it before so its
1376          * last_trans == cur_transid.
1377          * Or it won't be committed again onto disk after later
1378          * insert_dir_item()
1379          */
1380         if (!ret)
1381                 record_root_in_trans(trans, parent, 1);
1382         return ret;
1383 }
1384
1385 /*
1386  * new snapshots need to be created at a very specific time in the
1387  * transaction commit.  This does the actual creation.
1388  *
1389  * Note:
1390  * If the error which may affect the commitment of the current transaction
1391  * happens, we should return the error number. If the error which just affect
1392  * the creation of the pending snapshots, just return 0.
1393  */
1394 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1395                                    struct btrfs_pending_snapshot *pending)
1396 {
1397
1398         struct btrfs_fs_info *fs_info = trans->fs_info;
1399         struct btrfs_key key;
1400         struct btrfs_root_item *new_root_item;
1401         struct btrfs_root *tree_root = fs_info->tree_root;
1402         struct btrfs_root *root = pending->root;
1403         struct btrfs_root *parent_root;
1404         struct btrfs_block_rsv *rsv;
1405         struct inode *parent_inode;
1406         struct btrfs_path *path;
1407         struct btrfs_dir_item *dir_item;
1408         struct dentry *dentry;
1409         struct extent_buffer *tmp;
1410         struct extent_buffer *old;
1411         struct timespec64 cur_time;
1412         int ret = 0;
1413         u64 to_reserve = 0;
1414         u64 index = 0;
1415         u64 objectid;
1416         u64 root_flags;
1417         uuid_le new_uuid;
1418
1419         ASSERT(pending->path);
1420         path = pending->path;
1421
1422         ASSERT(pending->root_item);
1423         new_root_item = pending->root_item;
1424
1425         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1426         if (pending->error)
1427                 goto no_free_objectid;
1428
1429         /*
1430          * Make qgroup to skip current new snapshot's qgroupid, as it is
1431          * accounted by later btrfs_qgroup_inherit().
1432          */
1433         btrfs_set_skip_qgroup(trans, objectid);
1434
1435         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1436
1437         if (to_reserve > 0) {
1438                 pending->error = btrfs_block_rsv_add(root,
1439                                                      &pending->block_rsv,
1440                                                      to_reserve,
1441                                                      BTRFS_RESERVE_NO_FLUSH);
1442                 if (pending->error)
1443                         goto clear_skip_qgroup;
1444         }
1445
1446         key.objectid = objectid;
1447         key.offset = (u64)-1;
1448         key.type = BTRFS_ROOT_ITEM_KEY;
1449
1450         rsv = trans->block_rsv;
1451         trans->block_rsv = &pending->block_rsv;
1452         trans->bytes_reserved = trans->block_rsv->reserved;
1453         trace_btrfs_space_reservation(fs_info, "transaction",
1454                                       trans->transid,
1455                                       trans->bytes_reserved, 1);
1456         dentry = pending->dentry;
1457         parent_inode = pending->dir;
1458         parent_root = BTRFS_I(parent_inode)->root;
1459         record_root_in_trans(trans, parent_root, 0);
1460
1461         cur_time = current_time(parent_inode);
1462
1463         /*
1464          * insert the directory item
1465          */
1466         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1467         BUG_ON(ret); /* -ENOMEM */
1468
1469         /* check if there is a file/dir which has the same name. */
1470         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1471                                          btrfs_ino(BTRFS_I(parent_inode)),
1472                                          dentry->d_name.name,
1473                                          dentry->d_name.len, 0);
1474         if (dir_item != NULL && !IS_ERR(dir_item)) {
1475                 pending->error = -EEXIST;
1476                 goto dir_item_existed;
1477         } else if (IS_ERR(dir_item)) {
1478                 ret = PTR_ERR(dir_item);
1479                 btrfs_abort_transaction(trans, ret);
1480                 goto fail;
1481         }
1482         btrfs_release_path(path);
1483
1484         /*
1485          * pull in the delayed directory update
1486          * and the delayed inode item
1487          * otherwise we corrupt the FS during
1488          * snapshot
1489          */
1490         ret = btrfs_run_delayed_items(trans);
1491         if (ret) {      /* Transaction aborted */
1492                 btrfs_abort_transaction(trans, ret);
1493                 goto fail;
1494         }
1495
1496         record_root_in_trans(trans, root, 0);
1497         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1498         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1499         btrfs_check_and_init_root_item(new_root_item);
1500
1501         root_flags = btrfs_root_flags(new_root_item);
1502         if (pending->readonly)
1503                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1504         else
1505                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1506         btrfs_set_root_flags(new_root_item, root_flags);
1507
1508         btrfs_set_root_generation_v2(new_root_item,
1509                         trans->transid);
1510         uuid_le_gen(&new_uuid);
1511         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1512         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1513                         BTRFS_UUID_SIZE);
1514         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1515                 memset(new_root_item->received_uuid, 0,
1516                        sizeof(new_root_item->received_uuid));
1517                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1518                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1519                 btrfs_set_root_stransid(new_root_item, 0);
1520                 btrfs_set_root_rtransid(new_root_item, 0);
1521         }
1522         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1523         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1524         btrfs_set_root_otransid(new_root_item, trans->transid);
1525
1526         old = btrfs_lock_root_node(root);
1527         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1528         if (ret) {
1529                 btrfs_tree_unlock(old);
1530                 free_extent_buffer(old);
1531                 btrfs_abort_transaction(trans, ret);
1532                 goto fail;
1533         }
1534
1535         btrfs_set_lock_blocking_write(old);
1536
1537         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1538         /* clean up in any case */
1539         btrfs_tree_unlock(old);
1540         free_extent_buffer(old);
1541         if (ret) {
1542                 btrfs_abort_transaction(trans, ret);
1543                 goto fail;
1544         }
1545         /* see comments in should_cow_block() */
1546         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1547         smp_wmb();
1548
1549         btrfs_set_root_node(new_root_item, tmp);
1550         /* record when the snapshot was created in key.offset */
1551         key.offset = trans->transid;
1552         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1553         btrfs_tree_unlock(tmp);
1554         free_extent_buffer(tmp);
1555         if (ret) {
1556                 btrfs_abort_transaction(trans, ret);
1557                 goto fail;
1558         }
1559
1560         /*
1561          * insert root back/forward references
1562          */
1563         ret = btrfs_add_root_ref(trans, objectid,
1564                                  parent_root->root_key.objectid,
1565                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1566                                  dentry->d_name.name, dentry->d_name.len);
1567         if (ret) {
1568                 btrfs_abort_transaction(trans, ret);
1569                 goto fail;
1570         }
1571
1572         key.offset = (u64)-1;
1573         pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1574         if (IS_ERR(pending->snap)) {
1575                 ret = PTR_ERR(pending->snap);
1576                 btrfs_abort_transaction(trans, ret);
1577                 goto fail;
1578         }
1579
1580         ret = btrfs_reloc_post_snapshot(trans, pending);
1581         if (ret) {
1582                 btrfs_abort_transaction(trans, ret);
1583                 goto fail;
1584         }
1585
1586         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1587         if (ret) {
1588                 btrfs_abort_transaction(trans, ret);
1589                 goto fail;
1590         }
1591
1592         /*
1593          * Do special qgroup accounting for snapshot, as we do some qgroup
1594          * snapshot hack to do fast snapshot.
1595          * To co-operate with that hack, we do hack again.
1596          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1597          */
1598         ret = qgroup_account_snapshot(trans, root, parent_root,
1599                                       pending->inherit, objectid);
1600         if (ret < 0)
1601                 goto fail;
1602
1603         ret = btrfs_insert_dir_item(trans, dentry->d_name.name,
1604                                     dentry->d_name.len, BTRFS_I(parent_inode),
1605                                     &key, BTRFS_FT_DIR, index);
1606         /* We have check then name at the beginning, so it is impossible. */
1607         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1608         if (ret) {
1609                 btrfs_abort_transaction(trans, ret);
1610                 goto fail;
1611         }
1612
1613         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1614                                          dentry->d_name.len * 2);
1615         parent_inode->i_mtime = parent_inode->i_ctime =
1616                 current_time(parent_inode);
1617         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1618         if (ret) {
1619                 btrfs_abort_transaction(trans, ret);
1620                 goto fail;
1621         }
1622         ret = btrfs_uuid_tree_add(trans, new_uuid.b, BTRFS_UUID_KEY_SUBVOL,
1623                                   objectid);
1624         if (ret) {
1625                 btrfs_abort_transaction(trans, ret);
1626                 goto fail;
1627         }
1628         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1629                 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1630                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1631                                           objectid);
1632                 if (ret && ret != -EEXIST) {
1633                         btrfs_abort_transaction(trans, ret);
1634                         goto fail;
1635                 }
1636         }
1637
1638         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1639         if (ret) {
1640                 btrfs_abort_transaction(trans, ret);
1641                 goto fail;
1642         }
1643
1644 fail:
1645         pending->error = ret;
1646 dir_item_existed:
1647         trans->block_rsv = rsv;
1648         trans->bytes_reserved = 0;
1649 clear_skip_qgroup:
1650         btrfs_clear_skip_qgroup(trans);
1651 no_free_objectid:
1652         kfree(new_root_item);
1653         pending->root_item = NULL;
1654         btrfs_free_path(path);
1655         pending->path = NULL;
1656
1657         return ret;
1658 }
1659
1660 /*
1661  * create all the snapshots we've scheduled for creation
1662  */
1663 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1664 {
1665         struct btrfs_pending_snapshot *pending, *next;
1666         struct list_head *head = &trans->transaction->pending_snapshots;
1667         int ret = 0;
1668
1669         list_for_each_entry_safe(pending, next, head, list) {
1670                 list_del(&pending->list);
1671                 ret = create_pending_snapshot(trans, pending);
1672                 if (ret)
1673                         break;
1674         }
1675         return ret;
1676 }
1677
1678 static void update_super_roots(struct btrfs_fs_info *fs_info)
1679 {
1680         struct btrfs_root_item *root_item;
1681         struct btrfs_super_block *super;
1682
1683         super = fs_info->super_copy;
1684
1685         root_item = &fs_info->chunk_root->root_item;
1686         super->chunk_root = root_item->bytenr;
1687         super->chunk_root_generation = root_item->generation;
1688         super->chunk_root_level = root_item->level;
1689
1690         root_item = &fs_info->tree_root->root_item;
1691         super->root = root_item->bytenr;
1692         super->generation = root_item->generation;
1693         super->root_level = root_item->level;
1694         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1695                 super->cache_generation = root_item->generation;
1696         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1697                 super->uuid_tree_generation = root_item->generation;
1698 }
1699
1700 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1701 {
1702         struct btrfs_transaction *trans;
1703         int ret = 0;
1704
1705         spin_lock(&info->trans_lock);
1706         trans = info->running_transaction;
1707         if (trans)
1708                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1709         spin_unlock(&info->trans_lock);
1710         return ret;
1711 }
1712
1713 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1714 {
1715         struct btrfs_transaction *trans;
1716         int ret = 0;
1717
1718         spin_lock(&info->trans_lock);
1719         trans = info->running_transaction;
1720         if (trans)
1721                 ret = is_transaction_blocked(trans);
1722         spin_unlock(&info->trans_lock);
1723         return ret;
1724 }
1725
1726 /*
1727  * wait for the current transaction commit to start and block subsequent
1728  * transaction joins
1729  */
1730 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1731                                             struct btrfs_transaction *trans)
1732 {
1733         wait_event(fs_info->transaction_blocked_wait,
1734                    trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1735 }
1736
1737 /*
1738  * wait for the current transaction to start and then become unblocked.
1739  * caller holds ref.
1740  */
1741 static void wait_current_trans_commit_start_and_unblock(
1742                                         struct btrfs_fs_info *fs_info,
1743                                         struct btrfs_transaction *trans)
1744 {
1745         wait_event(fs_info->transaction_wait,
1746                    trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1747 }
1748
1749 /*
1750  * commit transactions asynchronously. once btrfs_commit_transaction_async
1751  * returns, any subsequent transaction will not be allowed to join.
1752  */
1753 struct btrfs_async_commit {
1754         struct btrfs_trans_handle *newtrans;
1755         struct work_struct work;
1756 };
1757
1758 static void do_async_commit(struct work_struct *work)
1759 {
1760         struct btrfs_async_commit *ac =
1761                 container_of(work, struct btrfs_async_commit, work);
1762
1763         /*
1764          * We've got freeze protection passed with the transaction.
1765          * Tell lockdep about it.
1766          */
1767         if (ac->newtrans->type & __TRANS_FREEZABLE)
1768                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1769
1770         current->journal_info = ac->newtrans;
1771
1772         btrfs_commit_transaction(ac->newtrans);
1773         kfree(ac);
1774 }
1775
1776 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1777                                    int wait_for_unblock)
1778 {
1779         struct btrfs_fs_info *fs_info = trans->fs_info;
1780         struct btrfs_async_commit *ac;
1781         struct btrfs_transaction *cur_trans;
1782
1783         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1784         if (!ac)
1785                 return -ENOMEM;
1786
1787         INIT_WORK(&ac->work, do_async_commit);
1788         ac->newtrans = btrfs_join_transaction(trans->root);
1789         if (IS_ERR(ac->newtrans)) {
1790                 int err = PTR_ERR(ac->newtrans);
1791                 kfree(ac);
1792                 return err;
1793         }
1794
1795         /* take transaction reference */
1796         cur_trans = trans->transaction;
1797         refcount_inc(&cur_trans->use_count);
1798
1799         btrfs_end_transaction(trans);
1800
1801         /*
1802          * Tell lockdep we've released the freeze rwsem, since the
1803          * async commit thread will be the one to unlock it.
1804          */
1805         if (ac->newtrans->type & __TRANS_FREEZABLE)
1806                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1807
1808         schedule_work(&ac->work);
1809
1810         /* wait for transaction to start and unblock */
1811         if (wait_for_unblock)
1812                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1813         else
1814                 wait_current_trans_commit_start(fs_info, cur_trans);
1815
1816         if (current->journal_info == trans)
1817                 current->journal_info = NULL;
1818
1819         btrfs_put_transaction(cur_trans);
1820         return 0;
1821 }
1822
1823
1824 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1825 {
1826         struct btrfs_fs_info *fs_info = trans->fs_info;
1827         struct btrfs_transaction *cur_trans = trans->transaction;
1828
1829         WARN_ON(refcount_read(&trans->use_count) > 1);
1830
1831         btrfs_abort_transaction(trans, err);
1832
1833         spin_lock(&fs_info->trans_lock);
1834
1835         /*
1836          * If the transaction is removed from the list, it means this
1837          * transaction has been committed successfully, so it is impossible
1838          * to call the cleanup function.
1839          */
1840         BUG_ON(list_empty(&cur_trans->list));
1841
1842         list_del_init(&cur_trans->list);
1843         if (cur_trans == fs_info->running_transaction) {
1844                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1845                 spin_unlock(&fs_info->trans_lock);
1846                 wait_event(cur_trans->writer_wait,
1847                            atomic_read(&cur_trans->num_writers) == 1);
1848
1849                 spin_lock(&fs_info->trans_lock);
1850         }
1851         spin_unlock(&fs_info->trans_lock);
1852
1853         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1854
1855         spin_lock(&fs_info->trans_lock);
1856         if (cur_trans == fs_info->running_transaction)
1857                 fs_info->running_transaction = NULL;
1858         spin_unlock(&fs_info->trans_lock);
1859
1860         if (trans->type & __TRANS_FREEZABLE)
1861                 sb_end_intwrite(fs_info->sb);
1862         btrfs_put_transaction(cur_trans);
1863         btrfs_put_transaction(cur_trans);
1864
1865         trace_btrfs_transaction_commit(trans->root);
1866
1867         if (current->journal_info == trans)
1868                 current->journal_info = NULL;
1869         btrfs_scrub_cancel(fs_info);
1870
1871         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1872 }
1873
1874 /*
1875  * Release reserved delayed ref space of all pending block groups of the
1876  * transaction and remove them from the list
1877  */
1878 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
1879 {
1880        struct btrfs_fs_info *fs_info = trans->fs_info;
1881        struct btrfs_block_group_cache *block_group, *tmp;
1882
1883        list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
1884                btrfs_delayed_refs_rsv_release(fs_info, 1);
1885                list_del_init(&block_group->bg_list);
1886        }
1887 }
1888
1889 static inline int btrfs_start_delalloc_flush(struct btrfs_trans_handle *trans)
1890 {
1891         struct btrfs_fs_info *fs_info = trans->fs_info;
1892
1893         /*
1894          * We use writeback_inodes_sb here because if we used
1895          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1896          * Currently are holding the fs freeze lock, if we do an async flush
1897          * we'll do btrfs_join_transaction() and deadlock because we need to
1898          * wait for the fs freeze lock.  Using the direct flushing we benefit
1899          * from already being in a transaction and our join_transaction doesn't
1900          * have to re-take the fs freeze lock.
1901          */
1902         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT)) {
1903                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1904         } else {
1905                 struct btrfs_pending_snapshot *pending;
1906                 struct list_head *head = &trans->transaction->pending_snapshots;
1907
1908                 /*
1909                  * Flush dellaloc for any root that is going to be snapshotted.
1910                  * This is done to avoid a corrupted version of files, in the
1911                  * snapshots, that had both buffered and direct IO writes (even
1912                  * if they were done sequentially) due to an unordered update of
1913                  * the inode's size on disk.
1914                  */
1915                 list_for_each_entry(pending, head, list) {
1916                         int ret;
1917
1918                         ret = btrfs_start_delalloc_snapshot(pending->root);
1919                         if (ret)
1920                                 return ret;
1921                 }
1922         }
1923         return 0;
1924 }
1925
1926 static inline void btrfs_wait_delalloc_flush(struct btrfs_trans_handle *trans)
1927 {
1928         struct btrfs_fs_info *fs_info = trans->fs_info;
1929
1930         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT)) {
1931                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1932         } else {
1933                 struct btrfs_pending_snapshot *pending;
1934                 struct list_head *head = &trans->transaction->pending_snapshots;
1935
1936                 /*
1937                  * Wait for any dellaloc that we started previously for the roots
1938                  * that are going to be snapshotted. This is to avoid a corrupted
1939                  * version of files in the snapshots that had both buffered and
1940                  * direct IO writes (even if they were done sequentially).
1941                  */
1942                 list_for_each_entry(pending, head, list)
1943                         btrfs_wait_ordered_extents(pending->root,
1944                                                    U64_MAX, 0, U64_MAX);
1945         }
1946 }
1947
1948 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1949 {
1950         struct btrfs_fs_info *fs_info = trans->fs_info;
1951         struct btrfs_transaction *cur_trans = trans->transaction;
1952         struct btrfs_transaction *prev_trans = NULL;
1953         int ret;
1954
1955         /* Stop the commit early if ->aborted is set */
1956         if (unlikely(READ_ONCE(cur_trans->aborted))) {
1957                 ret = cur_trans->aborted;
1958                 btrfs_end_transaction(trans);
1959                 return ret;
1960         }
1961
1962         btrfs_trans_release_metadata(trans);
1963         trans->block_rsv = NULL;
1964
1965         /* make a pass through all the delayed refs we have so far
1966          * any runnings procs may add more while we are here
1967          */
1968         ret = btrfs_run_delayed_refs(trans, 0);
1969         if (ret) {
1970                 btrfs_end_transaction(trans);
1971                 return ret;
1972         }
1973
1974         cur_trans = trans->transaction;
1975
1976         /*
1977          * set the flushing flag so procs in this transaction have to
1978          * start sending their work down.
1979          */
1980         cur_trans->delayed_refs.flushing = 1;
1981         smp_wmb();
1982
1983         btrfs_create_pending_block_groups(trans);
1984
1985         ret = btrfs_run_delayed_refs(trans, 0);
1986         if (ret) {
1987                 btrfs_end_transaction(trans);
1988                 return ret;
1989         }
1990
1991         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1992                 int run_it = 0;
1993
1994                 /* this mutex is also taken before trying to set
1995                  * block groups readonly.  We need to make sure
1996                  * that nobody has set a block group readonly
1997                  * after a extents from that block group have been
1998                  * allocated for cache files.  btrfs_set_block_group_ro
1999                  * will wait for the transaction to commit if it
2000                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2001                  *
2002                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2003                  * only one process starts all the block group IO.  It wouldn't
2004                  * hurt to have more than one go through, but there's no
2005                  * real advantage to it either.
2006                  */
2007                 mutex_lock(&fs_info->ro_block_group_mutex);
2008                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2009                                       &cur_trans->flags))
2010                         run_it = 1;
2011                 mutex_unlock(&fs_info->ro_block_group_mutex);
2012
2013                 if (run_it) {
2014                         ret = btrfs_start_dirty_block_groups(trans);
2015                         if (ret) {
2016                                 btrfs_end_transaction(trans);
2017                                 return ret;
2018                         }
2019                 }
2020         }
2021
2022         spin_lock(&fs_info->trans_lock);
2023         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2024                 spin_unlock(&fs_info->trans_lock);
2025                 refcount_inc(&cur_trans->use_count);
2026                 ret = btrfs_end_transaction(trans);
2027
2028                 wait_for_commit(cur_trans);
2029
2030                 if (unlikely(cur_trans->aborted))
2031                         ret = cur_trans->aborted;
2032
2033                 btrfs_put_transaction(cur_trans);
2034
2035                 return ret;
2036         }
2037
2038         cur_trans->state = TRANS_STATE_COMMIT_START;
2039         wake_up(&fs_info->transaction_blocked_wait);
2040
2041         if (cur_trans->list.prev != &fs_info->trans_list) {
2042                 prev_trans = list_entry(cur_trans->list.prev,
2043                                         struct btrfs_transaction, list);
2044                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2045                         refcount_inc(&prev_trans->use_count);
2046                         spin_unlock(&fs_info->trans_lock);
2047
2048                         wait_for_commit(prev_trans);
2049                         ret = prev_trans->aborted;
2050
2051                         btrfs_put_transaction(prev_trans);
2052                         if (ret)
2053                                 goto cleanup_transaction;
2054                 } else {
2055                         spin_unlock(&fs_info->trans_lock);
2056                 }
2057         } else {
2058                 spin_unlock(&fs_info->trans_lock);
2059         }
2060
2061         extwriter_counter_dec(cur_trans, trans->type);
2062
2063         ret = btrfs_start_delalloc_flush(trans);
2064         if (ret)
2065                 goto cleanup_transaction;
2066
2067         ret = btrfs_run_delayed_items(trans);
2068         if (ret)
2069                 goto cleanup_transaction;
2070
2071         wait_event(cur_trans->writer_wait,
2072                    extwriter_counter_read(cur_trans) == 0);
2073
2074         /* some pending stuffs might be added after the previous flush. */
2075         ret = btrfs_run_delayed_items(trans);
2076         if (ret)
2077                 goto cleanup_transaction;
2078
2079         btrfs_wait_delalloc_flush(trans);
2080
2081         btrfs_scrub_pause(fs_info);
2082         /*
2083          * Ok now we need to make sure to block out any other joins while we
2084          * commit the transaction.  We could have started a join before setting
2085          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2086          */
2087         spin_lock(&fs_info->trans_lock);
2088         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2089         spin_unlock(&fs_info->trans_lock);
2090         wait_event(cur_trans->writer_wait,
2091                    atomic_read(&cur_trans->num_writers) == 1);
2092
2093         /* ->aborted might be set after the previous check, so check it */
2094         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2095                 ret = cur_trans->aborted;
2096                 goto scrub_continue;
2097         }
2098         /*
2099          * the reloc mutex makes sure that we stop
2100          * the balancing code from coming in and moving
2101          * extents around in the middle of the commit
2102          */
2103         mutex_lock(&fs_info->reloc_mutex);
2104
2105         /*
2106          * We needn't worry about the delayed items because we will
2107          * deal with them in create_pending_snapshot(), which is the
2108          * core function of the snapshot creation.
2109          */
2110         ret = create_pending_snapshots(trans);
2111         if (ret) {
2112                 mutex_unlock(&fs_info->reloc_mutex);
2113                 goto scrub_continue;
2114         }
2115
2116         /*
2117          * We insert the dir indexes of the snapshots and update the inode
2118          * of the snapshots' parents after the snapshot creation, so there
2119          * are some delayed items which are not dealt with. Now deal with
2120          * them.
2121          *
2122          * We needn't worry that this operation will corrupt the snapshots,
2123          * because all the tree which are snapshoted will be forced to COW
2124          * the nodes and leaves.
2125          */
2126         ret = btrfs_run_delayed_items(trans);
2127         if (ret) {
2128                 mutex_unlock(&fs_info->reloc_mutex);
2129                 goto scrub_continue;
2130         }
2131
2132         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2133         if (ret) {
2134                 mutex_unlock(&fs_info->reloc_mutex);
2135                 goto scrub_continue;
2136         }
2137
2138         /*
2139          * make sure none of the code above managed to slip in a
2140          * delayed item
2141          */
2142         btrfs_assert_delayed_root_empty(fs_info);
2143
2144         WARN_ON(cur_trans != trans->transaction);
2145
2146         /* btrfs_commit_tree_roots is responsible for getting the
2147          * various roots consistent with each other.  Every pointer
2148          * in the tree of tree roots has to point to the most up to date
2149          * root for every subvolume and other tree.  So, we have to keep
2150          * the tree logging code from jumping in and changing any
2151          * of the trees.
2152          *
2153          * At this point in the commit, there can't be any tree-log
2154          * writers, but a little lower down we drop the trans mutex
2155          * and let new people in.  By holding the tree_log_mutex
2156          * from now until after the super is written, we avoid races
2157          * with the tree-log code.
2158          */
2159         mutex_lock(&fs_info->tree_log_mutex);
2160
2161         ret = commit_fs_roots(trans);
2162         if (ret) {
2163                 mutex_unlock(&fs_info->tree_log_mutex);
2164                 mutex_unlock(&fs_info->reloc_mutex);
2165                 goto scrub_continue;
2166         }
2167
2168         /*
2169          * Since the transaction is done, we can apply the pending changes
2170          * before the next transaction.
2171          */
2172         btrfs_apply_pending_changes(fs_info);
2173
2174         /* commit_fs_roots gets rid of all the tree log roots, it is now
2175          * safe to free the root of tree log roots
2176          */
2177         btrfs_free_log_root_tree(trans, fs_info);
2178
2179         /*
2180          * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2181          * new delayed refs. Must handle them or qgroup can be wrong.
2182          */
2183         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2184         if (ret) {
2185                 mutex_unlock(&fs_info->tree_log_mutex);
2186                 mutex_unlock(&fs_info->reloc_mutex);
2187                 goto scrub_continue;
2188         }
2189
2190         /*
2191          * Since fs roots are all committed, we can get a quite accurate
2192          * new_roots. So let's do quota accounting.
2193          */
2194         ret = btrfs_qgroup_account_extents(trans);
2195         if (ret < 0) {
2196                 mutex_unlock(&fs_info->tree_log_mutex);
2197                 mutex_unlock(&fs_info->reloc_mutex);
2198                 goto scrub_continue;
2199         }
2200
2201         ret = commit_cowonly_roots(trans);
2202         if (ret) {
2203                 mutex_unlock(&fs_info->tree_log_mutex);
2204                 mutex_unlock(&fs_info->reloc_mutex);
2205                 goto scrub_continue;
2206         }
2207
2208         /*
2209          * The tasks which save the space cache and inode cache may also
2210          * update ->aborted, check it.
2211          */
2212         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2213                 ret = cur_trans->aborted;
2214                 mutex_unlock(&fs_info->tree_log_mutex);
2215                 mutex_unlock(&fs_info->reloc_mutex);
2216                 goto scrub_continue;
2217         }
2218
2219         btrfs_prepare_extent_commit(fs_info);
2220
2221         cur_trans = fs_info->running_transaction;
2222
2223         btrfs_set_root_node(&fs_info->tree_root->root_item,
2224                             fs_info->tree_root->node);
2225         list_add_tail(&fs_info->tree_root->dirty_list,
2226                       &cur_trans->switch_commits);
2227
2228         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2229                             fs_info->chunk_root->node);
2230         list_add_tail(&fs_info->chunk_root->dirty_list,
2231                       &cur_trans->switch_commits);
2232
2233         switch_commit_roots(cur_trans);
2234
2235         ASSERT(list_empty(&cur_trans->dirty_bgs));
2236         ASSERT(list_empty(&cur_trans->io_bgs));
2237         update_super_roots(fs_info);
2238
2239         btrfs_set_super_log_root(fs_info->super_copy, 0);
2240         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2241         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2242                sizeof(*fs_info->super_copy));
2243
2244         btrfs_update_commit_device_size(fs_info);
2245         btrfs_update_commit_device_bytes_used(cur_trans);
2246
2247         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2248         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2249
2250         btrfs_trans_release_chunk_metadata(trans);
2251
2252         spin_lock(&fs_info->trans_lock);
2253         cur_trans->state = TRANS_STATE_UNBLOCKED;
2254         fs_info->running_transaction = NULL;
2255         spin_unlock(&fs_info->trans_lock);
2256         mutex_unlock(&fs_info->reloc_mutex);
2257
2258         wake_up(&fs_info->transaction_wait);
2259
2260         ret = btrfs_write_and_wait_transaction(trans);
2261         if (ret) {
2262                 btrfs_handle_fs_error(fs_info, ret,
2263                                       "Error while writing out transaction");
2264                 mutex_unlock(&fs_info->tree_log_mutex);
2265                 goto scrub_continue;
2266         }
2267
2268         ret = write_all_supers(fs_info, 0);
2269         /*
2270          * the super is written, we can safely allow the tree-loggers
2271          * to go about their business
2272          */
2273         mutex_unlock(&fs_info->tree_log_mutex);
2274         if (ret)
2275                 goto scrub_continue;
2276
2277         btrfs_finish_extent_commit(trans);
2278
2279         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2280                 btrfs_clear_space_info_full(fs_info);
2281
2282         fs_info->last_trans_committed = cur_trans->transid;
2283         /*
2284          * We needn't acquire the lock here because there is no other task
2285          * which can change it.
2286          */
2287         cur_trans->state = TRANS_STATE_COMPLETED;
2288         wake_up(&cur_trans->commit_wait);
2289         clear_bit(BTRFS_FS_NEED_ASYNC_COMMIT, &fs_info->flags);
2290
2291         spin_lock(&fs_info->trans_lock);
2292         list_del_init(&cur_trans->list);
2293         spin_unlock(&fs_info->trans_lock);
2294
2295         btrfs_put_transaction(cur_trans);
2296         btrfs_put_transaction(cur_trans);
2297
2298         if (trans->type & __TRANS_FREEZABLE)
2299                 sb_end_intwrite(fs_info->sb);
2300
2301         trace_btrfs_transaction_commit(trans->root);
2302
2303         btrfs_scrub_continue(fs_info);
2304
2305         if (current->journal_info == trans)
2306                 current->journal_info = NULL;
2307
2308         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2309
2310         return ret;
2311
2312 scrub_continue:
2313         btrfs_scrub_continue(fs_info);
2314 cleanup_transaction:
2315         btrfs_trans_release_metadata(trans);
2316         btrfs_cleanup_pending_block_groups(trans);
2317         btrfs_trans_release_chunk_metadata(trans);
2318         trans->block_rsv = NULL;
2319         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2320         if (current->journal_info == trans)
2321                 current->journal_info = NULL;
2322         cleanup_transaction(trans, ret);
2323
2324         return ret;
2325 }
2326
2327 /*
2328  * return < 0 if error
2329  * 0 if there are no more dead_roots at the time of call
2330  * 1 there are more to be processed, call me again
2331  *
2332  * The return value indicates there are certainly more snapshots to delete, but
2333  * if there comes a new one during processing, it may return 0. We don't mind,
2334  * because btrfs_commit_super will poke cleaner thread and it will process it a
2335  * few seconds later.
2336  */
2337 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2338 {
2339         int ret;
2340         struct btrfs_fs_info *fs_info = root->fs_info;
2341
2342         spin_lock(&fs_info->trans_lock);
2343         if (list_empty(&fs_info->dead_roots)) {
2344                 spin_unlock(&fs_info->trans_lock);
2345                 return 0;
2346         }
2347         root = list_first_entry(&fs_info->dead_roots,
2348                         struct btrfs_root, root_list);
2349         list_del_init(&root->root_list);
2350         spin_unlock(&fs_info->trans_lock);
2351
2352         btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2353
2354         btrfs_kill_all_delayed_nodes(root);
2355
2356         if (btrfs_header_backref_rev(root->node) <
2357                         BTRFS_MIXED_BACKREF_REV)
2358                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2359         else
2360                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2361
2362         return (ret < 0) ? 0 : 1;
2363 }
2364
2365 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2366 {
2367         unsigned long prev;
2368         unsigned long bit;
2369
2370         prev = xchg(&fs_info->pending_changes, 0);
2371         if (!prev)
2372                 return;
2373
2374         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2375         if (prev & bit)
2376                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2377         prev &= ~bit;
2378
2379         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2380         if (prev & bit)
2381                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2382         prev &= ~bit;
2383
2384         bit = 1 << BTRFS_PENDING_COMMIT;
2385         if (prev & bit)
2386                 btrfs_debug(fs_info, "pending commit done");
2387         prev &= ~bit;
2388
2389         if (prev)
2390                 btrfs_warn(fs_info,
2391                         "unknown pending changes left 0x%lx, ignoring", prev);
2392 }